![Backup: Optimization Strategy 1
NS1 NS2
m1
m2
m3
m4
ES1 ES3
ES4ES2
TT
RC
RC
RC
m1: 0.12
m3 utility: 0
m2: 1.87
m4 utility: 0.98
NS1 NS2
m1
m2
m3
m4
ES1 ES3
ES4ES2
TT
RC
RC
RC
m1: 0.12
m3 utility: 0
m2: 2.4
m4 utility: 1.31
NS1 NS2
m1
m2
m3
m4
ES1 ES3
ES4
ES2
RC
RC
RC
RC
m1: 1.62
m3 utility: 0.38
m2: 1.83
m4 utility: 1.96
ES4
(a) The current solution; Cost=0.98
Message T C link SS /(BAG, Lmax ) iterations
m1 TT NS1 − NS2 [0.09] 14
m2 RC — (4, 125) 5
m3 TT ES1 − NS1 [1] 0
m3 TT NS1 − NS2 [1.3] 7
(b) Tabu list
18 / 20](https://image.slidesharecdn.com/main-180707144444/85/Traffic-Class-Assignment-for-Mixed-Criticality-Frames-in-TTEthernet-18-320.jpg)
Traffic Class Assignment for Mixed-Criticality Frames in TTEthernet
- 1. Traffic Class Assignment for Mixed-Criticality Frames in TTEthernet Voica Gavrilut¸ Paul Pop Technical University of Denmark Kongens Lyngby, Denmark
- 2. Outline Motivation Architecture and application models The TTEthernet protocol Problem formulation and motivational example Optimization strategy: Tabu Search and cost function Experimental results Summary, message and future work 2 / 20
- 3. Motivation Trend: From “federated” to “integrated” architectures, where distributed applications of different criticality share the same platform Mixed-criticality systems: integrate safety-critical, mission-critical and non-critical applications Our focus is on the mixed time-criticality: hard real-time, soft real-time and non critical (non real-time) Hard real-time: missing a deadline leads to failure Soft real-time: missing a deadline degrades the service Safety-critical communication protocols: Specialized protocols in each area (e.g., CAN, FlexRay, SAFEBus, ProfiNet) Trend: extending Ethernet Ethernet: low cost, high speed, but unsuitable for real-time & safety-critical systems Extensions: AFDX/ARINC 664p7, EtherCAT, FTT-Ethernet, TTEthernet (our focus) 3 / 20
- 4. TTEthernet Traffic Classes TTEthernet Standardized as SAE 6802 ARINC 664p7 compliant Developed and marketed by TTTech Computertechnik AG Used in several application areas: automotive, aerospace, industrial Multiple traffic classes support mixed-criticality requirements Time-Triggered (TT) Very low latency and jitter The frames are sent based on schedule tables; highest priority Rate-Constrained (RC) Compatible with ARINC 664p7; lower priority than TT Guaranteed bandwidth via a “Bandwidth Allocation Gap” (BAG) Bounded worst-case end-to-end latency Best-Effort (BE) Standard Ethernet frame No timing guarantees; lowest priority Our problem: how to assign the traffic classes to mixed-criticality messages 4 / 20
- 5. Architecture Model ES1ES1 ES2ES2 ES3ES3 ES4ES4 NS1NS1 NS2NS2 vl1 τ1 τ4 τ2 τ5 τ3 A1: τ1, τ2, τ3 highly crical A2: τ4, τ5 non-crical NS3NS3 vl2 Virtual Link: tree-like structure, with on sender and mulple receivers Virtual Links (VL) Emulate point-to-point connections and provide the separation required for messages of mixed-criticality Each message has a VL, and we assume that VL routing is given 5 / 20
- 6. Application Model Mixed-criticality messages HRT: periodic hard real-time messages with a hard deadline SRT: periodic or sporadic soft real-time messages with a utility function NC: aperiodic non-critical messages Message Source Destination(s) Size Period Deadline/ Utility m1 ∈ MHRT ES1 {ES3, ES4} 80 B 750 µs 200 µs m2 ∈ MSRT ES3 {ES2} 300 B 2 ms 1.4 ms/ see utility fig. m3 ∈ MNC ES2 {ES1, ES3} 1200 B - - 6 t utility 1.4 soft deadline Figure: Example utility(t) function for SRT messages 6 / 20
- 7. TTEthernet: TT and RC traffic TT Traffic TT frames are sent based on schedule tables and have the highest priority The schedules contain the time when TT frames are sent and received on the links RC Traffic RC frames are queued up at the outgoing ports, and have to wait for TT frames and other RC frames A “Traffic Regulator” assures that there is at most one frame sent during a BAG interval Lmax is the maximum size of a RC frame Traffic integration policies: Preemtion The transmission of lower priority message is interrupted and resumed after the integral transmission of the higher priority message *timely block The lower priority message transmission is postponed if it would interfere with the transmission of a scheduled higher priority message Shuffling The transmission of higher priority message is postponed until the lower priority messages sending is finished 7 / 20
- 8. Problem Formulation Given The architecture model; the TTEthernet cluster The application model M = MHRT ∪ MSRT including all the message properties Note: for each message we know its VL and the VL routing Determine The traffic class T C(mi ) for each message mi The BAG and Lmax for each RC message The sending schedule tables SS for each TT message Such that The HRT messages are schedulable and The total utility for SRT messages is maximized. 8 / 20
- 9. Motivational Example: Introduction NS1 NS2 vl1m1 m2 m3 m4 ES1 ES3 ES4ES2 HRT HRT SRT SRT (a) Example architecture model Msg. Size Period Deadline / (Utility) m1 ∈ MHRT 50 B 2 ms 1 ms m2 ∈ MHRT 62.5 B 3 ms 2 ms m3 ∈ MSRT 500 B 4 ms 1.5 ms / (max. 6; 0 at 2.6 ms) m4 ∈ MSRT 750 B 4 ms 2.5 ms / (max. 6; 0 at 4.1 ms) (b) Example application model t utility 6 1.5 t utility 6 2.5 (c) Example Utility Functions 9 / 20
- 10. Motivational Example NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 RC RC RC RC NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT TT RC RC m1:1.96 1 m3:2.53; 0.38 m2:1.83 2 m4:2.72; 3.94 m1:0.12 1 m3:3.27; 0 m2:0.15 2 m4:3.64; 1.31 (a) All messages are RC; m1 is not schedulable; total achieved utility is only 36% out of 12. NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 RC RC RC RC NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT TT RC RC m1:1.96 1 m3:2.53; 0.38 m2:1.83 2 m4:2.72; 3.94 m1:0.12 1 m3:3.27; 0 m2:0.15 2 m4:3.64; 1.31 (b) HRT messages are TT and SRT are RC. m1 and m2 are schedulable, but the total utility is only of 11%. NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC TT RC m1:0.12 1 m3:3.91; 0 m2:1.91 2 m4:2.96; 3.25 m1:0.12 1 m3:1.2; 6 max. m2:1.98 2 m4:2.88; 3.48 (c) HRT m2 is RC, SRT m3 is TT. HRT are schedulable, and the total utility is increased to 79%. m3 has a maximum utility. 10 / 20
- 11. Optimization Strategy: Tabu Search Tabu Search meta-heuristic Search heuristic Explores the search space using Design Transformations Maximizes the Cost Function Avoids revisiting recent solutions by labeling them as “tabu” Cost Function Cost(Ψ) = wpHRT · δHRT + mi ∈MSRT mi .utility(WCD(mi )) Degree of schedulability: δHRT = mi ∈MHRT min(0, mi .deadline − WCD(mi )) WCD is the worst-case end-to-end delay TT: given by the schedule table RC: determined using a trajectory approach-based analysis method Design Transformations Switch Traffic Class: switches the traffic class of a message Modify Schedule: advances or postpones a TT frame Modify VL: increases or decreases the BAG and Lmax 11 / 20
- 12. Experimental results Name No. HRT msgs. No. SRT msgs. SFS TCA %HRT sched. %SRT utility Running time (h:min) %HRT sched. %SRT utility tc1 9 11 44.44% 90.27% 00:50 100% 100% tc2 11 23 54.54% 85.07% 2:30 100% 99.63% tc3 17 28 47.06% 64.10% 3:45 100% 95.77% SAE 40 39 70.00% 81.72% 5:00 100% 94.61% orion 99 87 45.45% 78.80% 12:30 94.94% 98.68% Evaluated algorithms: Traffic Class Assignment (TCA): Our proposed Tabu Search optimization Straightforward Solution (SFS): all messages are RC, and BAG and Lmax are optimized 3 synthetic cases and 2 real-life The synthetic test cases have the same topology with an increasing number of messages SAE is the “SAE automotive communication” benchmark Orion is the “Orion Crew Exploration Vehicle” case study Implementation and hardware: Java programming language (JDK1.8) Intel Xeon E5-2665 at 2.4 GHz 12 / 20
- 13. Summary and message Summary Addressed mixed-criticality applications implemented over TTEthernet networks Problem: decide the traffic class of each message Solution: Tabu Search-based optimization strategy Message For mixed-criticality message it is not obvious what is the best traffic class We need tools to decide the assignment of traffic classes Future work Handle the fragmenting and packing of TT frames Consider that the traffic class is assigned per dataflow link and not per message Ongoing: comparing against an SMT-based solution 13 / 20
- 14. Discussion Advantages Disadvantages TT Can provide low latency and jitter There is a SMT-based schedules synthesizer that can handle large systems Has the most predictable behaviour due to the scheduled traffic Schedules are not flexible (difficult to add new messages) The SMT-based approach cannot take into account the RC traffic RC traffic is still used for legacy reasons Uses more bandwidth due to the integration policy RC There are methods to compute the WCD, so the latency can be bounded Uses less bandwidth Better suited for sporadic traffic; no wasted resources More flexible (easier to add new messages) Larger latency and jitter Requires complex analysis and optimization methods for bounded latency and resources utilization 14 / 20
- 15. Backup: TTEthernet: TT Example b b CPU P1,1 τ1 P1,2 τ2 B2,Tx B1,Tx TTS P1,3 P2,1 τ4 P2,2 τ3 P2,3 CPU FU B1,Rx B2,Rx ES1 ES2 NS2 NS3 FU TTR B1,Tx B2,Tx TTS NS1 SS f2 f3 f4 TT a c d e f g h i j k l m SR SS A1: τ1 à m1 à τ3, RC A2: τ2 à m2 à τ4, TT a c d e f g h i j k l m Packing message m2 into frame f2 Place f2 in buffer B1,Tx for transmission Send time specified in send schedule SS TTS sends f2 to NS1 f2 is sent on the dataflow link to NS1 The Filtering Unit (FU) checks the frame f2 Expected receive time specified in receive schedule SR TTR checks if f2 arrives according to schedule Place f2 in buffer B1,Tx for transmission Send time specified in send schedule SS FU checks f2 Store the frame into receive buffer B2,Rx Task τ4 reads f2 from buffer b 15 / 20
- 16. Backup: TTEthernet: RC Example CPU P1,1 τ1 P1,2 τ2 Q1,Tx Q2,Tx B2,Tx B1,Tx TR2 TR1 RCS TTS P1,3 P2,1 τ4 P2,2 τ3 P2,3 CPU FU Q1,Rx Q2,Rx B1,Rx B2,Rx ES1 ES2 NS2 NS3 FU TP TTR B1,Tx B2,Tx TTS NS1 SS f2 f3 f4 f1 RC TT QTx 1 2 3 4 5 6 7 8 9 10 11 12 13 SR SS 1 Packing message m1 into frame f1 2 Insert it in queue Q1,Tx 3 Traffic Regulator (TR) ensures bandwidth for each VL 4 RC scheduler RCmultiplexes frames coming from TRs 5 TTS transmits f1 when there is no TT traffic 6 f1 is sent on the dataflow link to NS1 7 FU checks the validity of the frame 8 Traffic Policing (TP) checks that f2 arrives according to the BAG 9 Copy f1 to outgoing queue QTx 10 Send f1 when there is no TT traffic 11 FU checks f1 12 Copy to receiving Q2,Rx 13 Task τ3 reads f1 from the queue A1: τ1 à m1 à τ3, RC A2: τ2 à m2 à τ4, TT 16 / 20
- 17. Backup: Optimization Strategy: Design Transformations Switch Traffic Class STC(mi ); switches the traffic class of a message: From RC to TT, uses an initial schedules generator From TT to RC, uses mi .period and mi .size to determine the vli parameters Modify Schedules MS(mi , postpone); affects only TT messages and postpones (when postpone = TRUE) or advances the schedules of a message, on all links, keeping the transmission sequence valid Modify VL BAG and Lmax MVL(mi , increase); affects only RC messages and doubles (when increase = TRUE) or halved the vli .BAG and vli .Lmax 17 / 20
- 18. Backup: Optimization Strategy 1 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 1.87 m4 utility: 0.98 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 2.4 m4 utility: 1.31 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4 ES2 RC RC RC RC m1: 1.62 m3 utility: 0.38 m2: 1.83 m4 utility: 1.96 ES4 (a) The current solution; Cost=0.98 Message T C link SS /(BAG, Lmax ) iterations m1 TT NS1 − NS2 [0.09] 14 m2 RC — (4, 125) 5 m3 TT ES1 − NS1 [1] 0 m3 TT NS1 − NS2 [1.3] 7 (b) Tabu list 18 / 20
- 19. Backup: Optimization Strategy 2 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 1.87 m4 utility: 0.98 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 2.4 m4 utility: 1.31 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4 ES2 RC RC RC RC m1: 1.62 m3 utility: 0.38 m2: 1.83 m4 utility: 1.96 ES4 (c) Modify RC VL: BAG and Lmax are doubled; Cost = −1.89; tabu NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 1.87 m4 utility: 0.98 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 2.4 m4 utility: 1.31 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4 ES2 RC RC RC RC m1: 1.62 m3 utility: 0.38 m2: 1.83 m4 utility: 1.96 ES4 (d) Switch Traffic Class of m1 from TT to RC; Cost = −2.62; non-tabu 19 / 20
- 20. Backup: Optimization Strategy 3 NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC TT RC m1: 0.15 m3 utility: 6 m2: 1.92 m4 utility: 3.48 ES4 NS1 NS2 m1 m2 m3 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 1.87 m4 utility: 0.98m4 (e) Switch Traffic Class of m3 from RC to TT; Cost = 9.48; non-tabu NS1 NS2 m1 m2 m3 m4 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.15 m3 utility: 6 m2: 1.92 m4 utility: 0.71 ES4 NS1 NS2 m1 m2 m3 ES1 ES3 ES4ES2 TT RC RC RC m1: 0.12 m3 utility: 0 m2: 1.87 m4 utility: 0.98m4 (f) Modify Schedule of m1 on ES1 − NS1 by postponing it with 0.04 ms; Cost = 0.98; non-tabu 20 / 20